[LU-4507] Server hangs and terrible performance - ZFS IOR

• Type: Bug
• Resolution: Cannot Reproduce
• Priority: Blocker
• Affects Version/s: Lustre 2.6.0
• Environment: Hyperion/LLNL
• Severity: 3
• Assignee: Isaac Huang (Inactive)
• Reporter: Cliff White (Inactive)

Description

For sometime now we have been observing terrible read performance when running ZFS IOR file-per-proccess. The system will see ~7 GB/s reading with ldiskfs, at higher client counts the ZFS read performance on this test will drop to ~400 MB/s which is roughly a single client level.
Observing the OSTs we typically see one or two of the 12 OSTs with a very high load, the rest idle. The busy OST with then timeout, frequently evict several clients, and move forward. Stack dumps and errors from two servers are attached. These tests are ongoing, please advise what further data needs to be collected.

Issue Links

• is related to: LU-4716 replay-ost-single test_5: stuck in dbuf_read->zio_wait